(release 20.x non-upstream patch) [LV] Add initial support for vector…

…izing literal struct return values (#82)

This patch adds initial support for vectorizing literal struct return
values. Currently, this is limited to the case where the struct is
homogeneous (all elements have the same type) and not packed. The users
of the call also must all be `extractvalue` instructions.

The intended use case for this is vectorizing intrinsics such as:

```
declare { float, float } @llvm.sincos.f32(float %x)
```

Mapping them to structure-returning library calls such as:

```
declare { <4 x float>, <4 x i32> } @Sleef_sincosf4_u10advsimd(<4 x float>)
```

Or their widened form (such as `@llvm.sincos.v4f32` in this case).

Implementing this required two main changes:

1. Supporting widening `extractvalue`
2. Adding support for vectorized struct types in LV
  * This is mostly limited to parts of the cost model and scalarization

Since the supported use case is narrow, the required changes are
relatively small.

---

Downstream issue: #87

llvm/include/llvm/Analysis/TargetTransformInfo.h

@@ -1473,6 +1473,14 @@ class TargetTransformInfo {
                                      TTI::TargetCostKind CostKind,
                                      unsigned Index = -1) const;
 
+  /* Downstream change: #87 (sincos vectorization)*/
+  /// \return The expected cost of aggregate inserts and extracts. This is
+  /// used when the instruction is not available; a typical use case is to
+  /// provision the cost of vectorization/scalarization in vectorizer passes.
+  InstructionCost getInsertExtractValueCost(unsigned Opcode,
+                                            TTI::TargetCostKind CostKind) const;
+  /* End downstream change: #87 */
+
   /// \return The cost of replication shuffle of \p VF elements typed \p EltTy
   /// \p ReplicationFactor times.
   ///
@@ -2205,6 +2213,11 @@ class TargetTransformInfo::Concept {
                             const APInt &DemandedDstElts,
                             TTI::TargetCostKind CostKind) = 0;
 
+  /* Downstream change: #87 (sincos vectorization)*/
+  virtual InstructionCost
+  getInsertExtractValueCost(unsigned Opcode, TTI::TargetCostKind CostKind) = 0;
+  /* End downstream change: #87 */
+
   virtual InstructionCost
   getMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
                   unsigned AddressSpace, TTI::TargetCostKind CostKind,
@@ -2926,6 +2939,13 @@ class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
     return Impl.getReplicationShuffleCost(EltTy, ReplicationFactor, VF,
                                           DemandedDstElts, CostKind);
   }
+  /* Downstream change: #87 (sincos vectorization)*/
+  InstructionCost
+  getInsertExtractValueCost(unsigned Opcode,
+                            TTI::TargetCostKind CostKind) override {
+    return Impl.getInsertExtractValueCost(Opcode, CostKind);
+  }
+  /* End downstream change: #87 */
   InstructionCost getMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
                                   unsigned AddressSpace,
                                   TTI::TargetCostKind CostKind,

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

@@ -745,6 +745,19 @@ class TargetTransformInfoImplBase {
     return 1;
   }
 
+  /* Downstream change: #87 (sincos vectorization)*/
+  InstructionCost
+  getInsertExtractValueCost(unsigned Opcode,
+                            TTI::TargetCostKind CostKind) const {
+    // Note: The `insertvalue` cost here is chosen to match the default case of
+    // getInstructionCost() -- as pior to adding this helper `insertvalue` was
+    // not handled.
+    if (Opcode == Instruction::InsertValue)
+      return CostKind == TTI::TCK_RecipThroughput ? -1 : TTI::TCC_Basic;
+    return TTI::TCC_Free;
+  }
+  /* End downstream change: #87 */
+
   InstructionCost getMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
                                   unsigned AddressSpace,
                                   TTI::TargetCostKind CostKind,
@@ -1296,9 +1309,13 @@ class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
     case Instruction::PHI:
     case Instruction::Switch:
       return TargetTTI->getCFInstrCost(Opcode, CostKind, I);
-    case Instruction::ExtractValue:
     case Instruction::Freeze:
       return TTI::TCC_Free;
+      /* Downstream change: #87 (sincos vectorization)*/
+    case Instruction::ExtractValue:
+    case Instruction::InsertValue:
+      return TargetTTI->getInsertExtractValueCost(Opcode, CostKind);
+      /* End downstream change: #87 */
     case Instruction::Alloca:
       if (cast<AllocaInst>(U)->isStaticAlloca())
         return TTI::TCC_Free;

llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

@@ -416,9 +416,9 @@ class LoopVectorizationLegality {
   /// has a vectorized variant available.
   bool hasVectorCallVariants() const { return VecCallVariantsFound; }
 
-  /// Returns true if there is at least one function call in the loop which
-  /// returns a struct type and needs to be vectorized.
-  bool hasStructVectorCall() const { return StructVecCallFound; }
+  /* Downstream change: #87 (sincos vectorization)*/
+  // Removed hasStructVectorCall()
+  /* End downstream change: #87 */
 
   unsigned getNumStores() const { return LAI->getNumStores(); }
   unsigned getNumLoads() const { return LAI->getNumLoads(); }
@@ -639,11 +639,9 @@ class LoopVectorizationLegality {
   /// the use of those function variants.
   bool VecCallVariantsFound = false;
 
-  /// If we find a call (to be vectorized) that returns a struct type, record
-  /// that so we can bail out until this is supported.
-  /// TODO: Remove this flag once vectorizing calls with struct returns is
-  /// supported.
-  bool StructVecCallFound = false;
+  /* Downstream change: #87 (sincos vectorization)*/
+  // Removed StructVecCallFound
+  /* End downstream change: #87 */
 
   /// Keep track of all the countable and uncountable exiting blocks if
   /// the exact backedge taken count is not computable.

llvm/lib/Analysis/TargetTransformInfo.cpp

@@ -1113,6 +1113,18 @@ TargetTransformInfo::getVectorInstrCost(const Instruction &I, Type *Val,
   return Cost;
 }
 
+/* Downstream change: #87 (sincos vectorization)*/
+InstructionCost TargetTransformInfo::getInsertExtractValueCost(
+    unsigned Opcode, TTI::TargetCostKind CostKind) const {
+  assert((Opcode == Instruction::InsertValue ||
+          Opcode == Instruction::ExtractValue) &&
+         "Expecting Opcode to be insertvalue/extractvalue.");
+  InstructionCost Cost = TTIImpl->getInsertExtractValueCost(Opcode, CostKind);
+  assert(Cost >= 0 && "TTI should not produce negative costs!");
+  return Cost;
+}
+/* End downstream change: #87 */
+
 InstructionCost TargetTransformInfo::getReplicationShuffleCost(
     Type *EltTy, int ReplicationFactor, int VF, const APInt &DemandedDstElts,
     TTI::TargetCostKind CostKind) const {

llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp

@@ -954,24 +954,19 @@ bool LoopVectorizationLegality::canVectorizeInstrs() {
       if (CI && !VFDatabase::getMappings(*CI).empty())
         VecCallVariantsFound = true;
 
-      auto CanWidenInstructionTy = [this](Instruction const &Inst) {
+      /* Downstream change: #87 (sincos vectorization)*/
+      auto CanWidenInstructionTy = [](Instruction const &Inst) {
         Type *InstTy = Inst.getType();
         if (!isa<StructType>(InstTy))
           return canVectorizeTy(InstTy);
 
         // For now, we only recognize struct values returned from calls where
         // all users are extractvalue as vectorizable. All element types of the
         // struct must be types that can be widened.
-        if (isa<CallInst>(Inst) && canWidenCallReturnType(InstTy) &&
-            all_of(Inst.users(), IsaPred<ExtractValueInst>)) {
-          // TODO: Remove the `StructVecCallFound` flag once vectorizing calls
-          // with struct returns is supported.
-          StructVecCallFound = true;
-          return true;
-        }
-
-        return false;
+        return isa<CallInst>(Inst) && canWidenCallReturnType(InstTy) &&
+               all_of(Inst.users(), IsaPred<ExtractValueInst>);
       };
+      /* End downstream change: #87 */
 
       // Check that the instruction return type is vectorizable.
       // We can't vectorize casts from vector type to scalar type.

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -2350,7 +2350,11 @@ void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr,
                                                VPReplicateRecipe *RepRecipe,
                                                const VPLane &Lane,
                                                VPTransformState &State) {
-  assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
+  /* Downstream change: #87 (sincos vectorization)*/
+  assert((!Instr->getType()->isAggregateType() ||
+          canVectorizeTy(Instr->getType())) &&
+         "Expected vectorizable or non-aggregate type.");
+  /* End downstream change: #87 */
 
   // Does this instruction return a value ?
   bool IsVoidRetTy = Instr->getType()->isVoidTy();
@@ -2855,11 +2859,13 @@ LoopVectorizationCostModel::getVectorCallCost(CallInst *CI,
   return ScalarCallCost;
 }
 
-static Type *maybeVectorizeType(Type *Elt, ElementCount VF) {
-  if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy()))
-    return Elt;
-  return VectorType::get(Elt, VF);
+/* Downstream change: #87 (sincos vectorization)*/
+static Type *maybeVectorizeType(Type *Ty, ElementCount VF) {
+  if (VF.isScalar() || !canVectorizeTy(Ty))
+    return Ty;
+  return toVectorizedTy(Ty, VF);
 }
+/* End downstream change: #87 */
 
 InstructionCost
 LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,
@@ -3605,14 +3611,18 @@ void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) {
         }
       }
 
-      // ExtractValue instructions must be uniform, because the operands are
-      // known to be loop-invariant.
+      /* Downstream change: #87 (sincos vectorization)*/
       if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) {
-        assert(IsOutOfScope(EVI->getAggregateOperand()) &&
-               "Expected aggregate value to be loop invariant");
-        AddToWorklistIfAllowed(EVI);
-        continue;
+        if (IsOutOfScope(EVI->getAggregateOperand())) {
+          AddToWorklistIfAllowed(EVI);
+          continue;
+        }
+        // Only ExtractValue instructions where the aggregate value comes from a
+        // call are allowed to be non-uniform.
+        assert(isa<CallInst>(EVI->getAggregateOperand()) &&
+               "Expected aggregate value to be call return value");
       }
+      /* End downstream change: #87 */
 
       // If there's no pointer operand, there's nothing to do.
       auto *Ptr = getLoadStorePointerOperand(&I);
@@ -4492,8 +4502,8 @@ static bool willGenerateVectors(VPlan &Plan, ElementCount VF,
         llvm_unreachable("unhandled recipe");
       }
 
-      auto WillWiden = [&TTI, VF](Type *ScalarTy) {
-        Type *VectorTy = toVectorTy(ScalarTy, VF);
+      /* Downstream change: #87 (sincos vectorization)*/
+      auto WillGenerateTargetVectors = [&TTI, VF](Type *VectorTy) {
         unsigned NumLegalParts = TTI.getNumberOfParts(VectorTy);
         if (!NumLegalParts)
           return false;
@@ -4505,9 +4515,10 @@ static bool willGenerateVectors(VPlan &Plan, ElementCount VF,
           // explicitly ask TTI about the register class uses for each part.
           return NumLegalParts <= VF.getKnownMinValue();
         }
-        // Two or more parts that share a register - are vectorized.
+        // Two or more elements that share a register - are vectorized.
         return NumLegalParts < VF.getKnownMinValue();
       };
+      /* End downstream change: #87 */
 
       // If no def nor is a store, e.g., branches, continue - no value to check.
       if (R.getNumDefinedValues() == 0 &&
@@ -4524,8 +4535,11 @@ static bool willGenerateVectors(VPlan &Plan, ElementCount VF,
       Type *ScalarTy = TypeInfo.inferScalarType(ToCheck);
       if (!Visited.insert({ScalarTy}).second)
         continue;
-      if (WillWiden(ScalarTy))
+      Type *WideTy = toVectorizedTy(ScalarTy, VF);
+      /* Downstream change: #87 (sincos vectorization)*/
+      if (any_of(getContainedTypes(WideTy), WillGenerateTargetVectors))
         return true;
+      /* End downstream change: #87 */
     }
   }
 
@@ -5481,10 +5495,15 @@ InstructionCost LoopVectorizationCostModel::computePredInstDiscount(
     // Compute the scalarization overhead of needed insertelement instructions
     // and phi nodes.
     if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) {
-      ScalarCost += TTI.getScalarizationOverhead(
-          cast<VectorType>(toVectorTy(I->getType(), VF)),
-          APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ true,
-          /*Extract*/ false, CostKind);
+      /* Downstream change: #87 (sincos vectorization)*/
+      Type *WideTy = toVectorizedTy(I->getType(), VF);
+      for (Type *VectorTy : getContainedTypes(WideTy)) {
+        ScalarCost += TTI.getScalarizationOverhead(
+            cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getFixedValue()),
+            /*Insert=*/true,
+            /*Extract=*/false, CostKind);
+      }
+      /* End downstream change: #87 */
       ScalarCost +=
           VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind);
     }
@@ -5495,16 +5514,21 @@ InstructionCost LoopVectorizationCostModel::computePredInstDiscount(
     // overhead.
     for (Use &U : I->operands())
       if (auto *J = dyn_cast<Instruction>(U.get())) {
-        assert(VectorType::isValidElementType(J->getType()) &&
+        /* Downstream change: #87 (sincos vectorization)*/
+        assert(canVectorizeTy(J->getType()) &&
                "Instruction has non-scalar type");
         if (CanBeScalarized(J))
           Worklist.push_back(J);
         else if (needsExtract(J, VF)) {
-          ScalarCost += TTI.getScalarizationOverhead(
-              cast<VectorType>(toVectorTy(J->getType(), VF)),
-              APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false,
-              /*Extract*/ true, CostKind);
+          Type *WideTy = toVectorizedTy(J->getType(), VF);
+          for (Type *VectorTy : getContainedTypes(WideTy)) {
+            ScalarCost += TTI.getScalarizationOverhead(
+                cast<VectorType>(VectorTy),
+                APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false,
+                /*Extract*/ true, CostKind);
+          }
         }
+        /* End downstream change: #87 */
       }
 
     // Scale the total scalar cost by block probability.
@@ -5982,13 +6006,19 @@ LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I,
     return 0;
 
   InstructionCost Cost = 0;
-  Type *RetTy = toVectorTy(I->getType(), VF);
+  /* Downstream change: #87 (sincos vectorization)*/
+  Type *RetTy = toVectorizedTy(I->getType(), VF);
   if (!RetTy->isVoidTy() &&
-      (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore()))
-    Cost += TTI.getScalarizationOverhead(
-        cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()),
-        /*Insert*/ true,
-        /*Extract*/ false, CostKind);
+      (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) {
+
+    for (Type *VectorTy : getContainedTypes(RetTy)) {
+      Cost += TTI.getScalarizationOverhead(
+          cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getKnownMinValue()),
+          /*Insert=*/true,
+          /*Extract=*/false, CostKind);
+    }
+  }
+  /* End downstream change: #87 */
 
   // Some targets keep addresses scalar.
   if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing())
@@ -6246,9 +6276,11 @@ void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) {
 
       bool MaskRequired = Legal->isMaskRequired(CI);
       // Compute corresponding vector type for return value and arguments.
-      Type *RetTy = toVectorTy(ScalarRetTy, VF);
+      /* Downstream change: #87 (sincos vectorization)*/
+      Type *RetTy = toVectorizedTy(ScalarRetTy, VF);
       for (Type *ScalarTy : ScalarTys)
-        Tys.push_back(toVectorTy(ScalarTy, VF));
+        Tys.push_back(toVectorizedTy(ScalarTy, VF));
+      /* End downstream change: #87 */
 
       // An in-loop reduction using an fmuladd intrinsic is a special case;
       // we don't want the normal cost for that intrinsic.
@@ -6425,7 +6457,8 @@ LoopVectorizationCostModel::getInstructionCost(Instruction *I,
            HasSingleCopyAfterVectorization(I, VF));
     VectorTy = RetTy;
   } else
-    VectorTy = toVectorTy(RetTy, VF);
+    // Downstream change: #87 (sincos vectorization)
+    VectorTy = toVectorizedTy(RetTy, VF);
 
   if (VF.isVector() && VectorTy->isVectorTy() &&
       !TTI.getNumberOfParts(VectorTy))
@@ -8547,7 +8580,7 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
   case Instruction::Shl:
   case Instruction::Sub:
   case Instruction::Xor:
-  case Instruction::Freeze:
+  case Instruction::Freeze: {
     SmallVector<VPValue *> NewOps(Operands);
     if (Instruction::isBinaryOp(I->getOpcode())) {
       // The legacy cost model uses SCEV to check if some of the operands are
@@ -8572,6 +8605,18 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
       NewOps[1] = GetConstantViaSCEV(NewOps[1]);
     }
     return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()));
+  }
+  /* Downstream change: #87 (sincos vectorization)*/
+  case Instruction::ExtractValue: {
+    SmallVector<VPValue *> NewOps(Operands);
+    Type *I32Ty = IntegerType::getInt32Ty(I->getContext());
+    auto *EVI = cast<ExtractValueInst>(I);
+    assert(EVI->getNumIndices() == 1 && "Expected one extractvalue index");
+    unsigned Idx = EVI->getIndices()[0];
+    NewOps.push_back(Plan.getOrAddLiveIn(ConstantInt::get(I32Ty, Idx, false)));
+    return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()));
+  }
+    /* End downstream change: #87 */
   };
 }
 
@@ -9852,7 +9897,8 @@ void VPReplicateRecipe::execute(VPTransformState &State) {
             VectorType::get(UI->getType(), State.VF));
         State.set(this, Poison);
       }
-      State.packScalarIntoVectorValue(this, *State.Lane);
+      // Downstream change: #87 (sincos vectorization)
+      State.packScalarIntoVectorizedValue(this, *State.Lane);
     }
     return;
   }
@@ -10369,12 +10415,9 @@ bool LoopVectorizePass::processLoop(Loop *L) {
     return false;
   }
 
-  if (LVL.hasStructVectorCall()) {
-    reportVectorizationFailure("Auto-vectorization of calls that return struct "
-                               "types is not yet supported",
-                               "StructCallVectorizationUnsupported", ORE, L);
-    return false;
-  }
+  /* Downstream change: #87 (sincos vectorization)*/
+  // Remove StructCallVectorizationUnsupported failure.
+  /* End downstream change: #87 */
 
   // Entrance to the VPlan-native vectorization path. Outer loops are processed
   // here. They may require CFG and instruction level transformations before